discourse/plugins/discourse-ai/lib/completions/llm.rb

# frozen_string_literal: true

# A facade that abstracts multiple LLMs behind a single interface.
#
# Internally, it consists of the combination of a dialect and an endpoint.
# After receiving a prompt using our generic format, it translates it to
# the target model and routes the completion request through the correct gateway.
#
# Use the .proxy method to instantiate an object.
# It chooses the correct dialect and endpoint for the model you want to interact with.
#
# Tests of modules that perform LLM calls can use .with_prepared_responses to return canned responses
# instead of relying on WebMock stubs like we did in the past.
#
module DiscourseAi
  module Completions
    class Llm
      UNKNOWN_MODEL = Class.new(StandardError)

      class << self
        def presets
          LlmPresets.all
        end

        def provider_names
          providers = %w[
            aws_bedrock
            aws_bedrock_converse
            anthropic
            vllm
            hugging_face
            cohere
            open_ai
            google
            google_vertex_ai
            azure
            samba_nova
            mistral
            open_router
            groq
          ]
          if !Rails.env.production?
            providers << "fake"
            providers << "ollama"
          end

          providers
        end

        def tokenizer_names
          DiscourseAi::Tokenizer::BasicTokenizer.available_llm_tokenizers.map(&:name)
        end

        # Endpoint capabilities the admin UI needs to render provider forms.
        # Endpoints match on provider (and sometimes url), so a lightweight
        # probe is enough to resolve them.
        def provider_capabilities
          probe = Struct.new(:provider, :url)

          provider_names.each_with_object({}) do |provider, capabilities|
            endpoint =
              begin
                DiscourseAi::Completions::Endpoints::Base.endpoint_for(probe.new(provider, ""))
              rescue UNKNOWN_MODEL
                nil
              end

            capabilities[provider] = {
              requires_configured_url: endpoint ? endpoint.requires_configured_url? : true,
            }
          end
        end

        def valid_provider_models
          return @valid_provider_models if defined?(@valid_provider_models)

          valid_provider_models = []
          models_by_provider.each do |provider, models|
            valid_provider_models.concat(models.map { |model| "#{provider}:#{model}" })
          end
          @valid_provider_models = Set.new(valid_provider_models)
        end

        def with_prepared_responses(responses, llm: nil)
          @canned_response = DiscourseAi::Completions::Endpoints::CannedResponse.new(responses)
          @canned_llm = llm
          @prompts = []
          @prompt_options = []

          yield(@canned_response, llm, @prompts, @prompt_options)
        ensure
          # Don't leak prepared response if there's an exception.
          @canned_response = nil
          @canned_llm = nil
          @prompts = nil
        end

        def record_prompt(prompt, options)
          @prompts << prompt.dup if @prompts
          @prompt_options << options if @prompt_options
        end

        def prompt_options
          @prompt_options
        end

        def prompts
          @prompts
        end

        def proxy(model)
          llm_model =
            if model.is_a?(LlmModel)
              model
            elsif model.is_a?(Numeric)
              LlmModel.find_by(id: model)
            else
              model_name_without_prov = model.split(":").last.to_i

              LlmModel.find_by(id: model_name_without_prov)
            end

          raise UNKNOWN_MODEL if llm_model.nil?

          dialect_klass = DiscourseAi::Completions::Dialects::Dialect.dialect_for(llm_model)

          if @canned_response
            if @canned_llm && @canned_llm != model
              raise "Invalid call LLM call, expected #{@canned_llm} but got #{model}"
            end

            return new(dialect_klass, nil, llm_model, gateway: @canned_response)
          end

          gateway_klass = DiscourseAi::Completions::Endpoints::Base.endpoint_for(llm_model)

          new(dialect_klass, gateway_klass, llm_model)
        end
      end

      def initialize(dialect_klass, gateway_klass, llm_model, gateway: nil)
        @dialect_klass = dialect_klass
        @gateway_klass = gateway_klass
        @gateway = gateway
        @llm_model = llm_model
      end

      # @param generic_prompt { DiscourseAi::Completions::Prompt } - Our generic prompt object
      # @param user { User } - User requesting the summary.
      # @param temperature { Float - Optional } - The temperature to use for the completion.
      # @param top_p { Float - Optional } - The top_p to use for the completion.
      # @param max_tokens { Integer - Optional } - The maximum number of tokens to generate.
      # @param stop_sequences { Array<String> - Optional } - The stop sequences to use for the completion.
      # @param feature_name { String - Optional } - The feature name to use for the completion.
      # @param feature_context { Hash - Optional } - The feature context to use for the completion.
      # @param partial_tool_calls { Boolean - Optional } - If true, the completion will return partial tool calls.
      # @param output_thinking { Boolean - Optional } - If true, the completion will return the thinking output for thinking models.
      # @param thinking_effort { String - Optional } - Provider-agnostic per-call thinking effort override.
      # @param response_format { Hash - Optional } - JSON schema passed to the API as the desired structured output.
      # @param [Experimental] extra_model_params { Hash - Optional } - Other params that are not available accross models. e.g. response_format JSON schema.
      # @param execution_context { DiscourseAi::Completions::ExecutionContext - Optional } - Explicit per-call context for token tracking and audit logging.
      #
      # @param &on_partial_blk { Block - Optional } - The passed block will get called with the LLM partial response.
      #
      # @returns String | ToolCall - Completion result.
      # if multiple tools or a tool and a message come back, the result will be an array of ToolCall / String objects.
      #
      def generate(
        prompt,
        temperature: nil,
        top_p: nil,
        max_tokens: nil,
        stop_sequences: nil,
        user:,
        feature_name: nil,
        feature_context: nil,
        partial_tool_calls: false,
        output_thinking: false,
        response_format: nil,
        thinking_effort: nil,
        extra_model_params: nil,
        cancel_manager: nil,
        execution_context: nil,
        &partial_read_blk
      )
        self.class.record_prompt(
          prompt,
          {
            temperature: temperature,
            top_p: top_p,
            max_tokens: max_tokens,
            stop_sequences: stop_sequences,
            user: user,
            feature_name: feature_name,
            feature_context: feature_context,
            partial_tool_calls: partial_tool_calls,
            output_thinking: output_thinking,
            response_format: response_format,
            thinking_effort: thinking_effort,
            extra_model_params: extra_model_params,
          },
        )

        model_params = {
          max_tokens: max_tokens,
          stop_sequences: stop_sequences,
          thinking_effort: thinking_effort,
        }

        if SiteSetting.ai_llm_temperature_top_p_enabled
          model_params[:temperature] = temperature if temperature
          model_params[:top_p] = top_p if top_p
        end

        # internals expect symbolized keys, so we normalize here
        response_format =
          JSON.parse(response_format.to_json, symbolize_names: true) if response_format &&
          response_format.is_a?(Hash)

        model_params[:response_format] = response_format if response_format
        model_params.merge!(extra_model_params) if extra_model_params

        if prompt.is_a?(String)
          prompt =
            DiscourseAi::Completions::Prompt.new(
              "You are a helpful bot",
              messages: [{ type: :user, content: prompt }],
            )
        elsif prompt.is_a?(Array)
          prompt = DiscourseAi::Completions::Prompt.new(messages: prompt)
        end

        if !prompt.is_a?(DiscourseAi::Completions::Prompt)
          raise ArgumentError, "Prompt must be either a string, array, of Prompt object"
        end

        model_params.keys.each { |key| model_params.delete(key) if model_params[key].nil? }

        gateway = @gateway || gateway_klass.new(llm_model)
        model_params = gateway.prepare_model_params(model_params) if gateway.respond_to?(
          :prepare_model_params,
        )
        dialect = dialect_klass.new(prompt, llm_model, opts: model_params)

        gateway.perform_completion!(
          dialect,
          user,
          model_params,
          feature_name: feature_name,
          feature_context: feature_context,
          partial_tool_calls: partial_tool_calls,
          output_thinking: output_thinking,
          cancel_manager: cancel_manager,
          execution_context:,
          &partial_read_blk
        )
      end

      def max_prompt_tokens
        llm_model.max_prompt_tokens
      end

      def tokenizer
        llm_model.tokenizer_class
      end

      attr_reader :llm_model

      private

      attr_reader :dialect_klass, :gateway_klass
    end
  end
end